The market's own record, as it stood on every date.

Predicate RX harmonizes the public records that describe healthcare entities, preserves every version without overwrite, and proves each fact back to its source. Delivered as a one-way feed into your environment; methods, including audit, run on it there.

crossing one direction at a time
4connections

That is the Crosswalk. One permanent identity per entity, linking every agency's number for it, so a question asked at one corner can be answered from any other — and keeps being answerable when a new corner appears.

Harmonized. Preserved. Proved.

Each step is what made the next one possible. The order is the argument.

Three properties, in the order they depend on each other

Not three features. A chain — and each link is worthless without the one before it.

Meaning attached to a misattributed value is a precisely described wrong answer. Provenance on a number nobody can interpret proves the origin of something unusable. The dependency is what makes this a layer rather than three columns you could add to a spreadsheet.

One account, asked once

An account manager wants to know whether this is one customer or four. It is an ordinary question, asked constantly, and the public record answers it four different ways.

THE SAME ORGANISATION, AS EACH AGENCY HOLDS IT
AGENCYIDENTIFIERWHAT IT NAMES GRAIN
MedicareCCN a billing entityfacility, sometimes campus
TaxEIN a filerlegal entity, often the whole system
Registry340B ID a participantparent, with its own child sites
Provider directoryNPI a providerorganisation or individual
SecuritiesCIK a registrantcorporate parent
Five identifiers. None of them is a key to the others, and none of the agencies checks with the rest. Your CRM holds a sixth.

Account management

Counts it as four accounts — four addresses, four contacts.

Finance

Counts it as one — it settles against a single parent.

Market access

Counts it as one parent and six sites, from the registry.

Analytics

Counts it as whichever source the model happened to load.

Four answers, none of them careless. The identifiers genuinely do not match, and no agency publishes a key. Reconciling them by hand is a project each desk runs separately, gets a different answer from, and repeats next quarter.

Put one resolved record underneath, and the four answers become one.

Not because anyone changed their process — because they are now querying the same thing. That is the part of this that operates across an organisation rather than inside a single value.

And one value, expanded

The crosswalk settles which organisation. This settles what the number means — which is what lets a department use it without a project, and a model use it without an engineer in between.

// ONE FIELD AND ITS SOURCEPRINT™
fieldcharity_care_at_cost
value48,113,000 USD
grainone value per CCN per fiscal year
sourceCMS Form 2552-10, Worksheet S-10, Line 23, Column 1
verified againstCMS Pub. 15-2, chapter 40
periodFY2023, settled version
observed2026-08-14, capture 0f3a91c2
if blankreported-zero and not-filed are distinct, never merged
do not concludenot uncompensated care; excludes bad debt and unreimbursed Medicaid; do not sum with uncomp_care_total — they overlap

Two of those lines exist because the alternative is an error that looks correct. A blank that means "never filed" and a blank that means "zero" are opposite facts, and almost nothing distinguishes them. Two adjacent columns that overlap will be summed by someone, eventually, and the total will look entirely reasonable.

The analyst who knew that field was a cost rather than a charge used to catch it. That person was the error-correction layer, and automating the analysis removes them. What replaces them has no suspicion — only fluency.

Field names never change meaning. They are added to, never repurposed. So work built on the vocabulary keeps working across form revisions, and a schema change does not silently alter what a downstream model believes.

What you buy

An annual subscription to the records feed. Not a platform, not a seat, not a report — the record itself, delivered into your environment and yours to build on.

01 — THE FEED

The records, as captured

Every source we hold, crosswalked to one entity identity and Sourceprinted at field level. Daily where the source publishes daily; on release where it doesn't.

Not just the current state — every past state, back to our first capture of each source, so a question about a date is answered from the record as it stood on that date.

02 — DELIVERY

One way, into your environment

Pull from an endpoint, a read-only share, or a share straight into your warehouse. It lands where your own data already lives.

Nothing of yours comes back to us. We do not see your queries, your joins, your results, or which entities you looked at.

03 — THE LICENCE

Flat annual fee, unlimited internal use

One fee for the organisation. Unlimited seats, unlimited queries, unlimited internal systems reading it.

Metering seats would penalise exactly the adoption that makes it worth having — the second and third desk to pick it up are where it starts paying for itself.

04 — WHAT YOU BUILD ON IT

Whatever the question is

Merge it with your CRM, your claims, your contracts, your forecast. Build internal tools, models and agents directly on it.

Run the published methods — Audit, Check, Watch — or write your own. They are open source and they run on your side, not ours.

05 — WHAT IS NOT IN IT

No analysis, no scores, no findings

The feed carries declared facts — what market parties stated about themselves in public filings — and nothing we concluded from them. No risk scores, no rankings, no recommendations.

That is deliberate. A record that takes a position is a record the other side can attack. This one states only what was filed, and when.

06 — WHY IT COMPOUNDS

The history cannot be bought later

The sources overwrite themselves. A party starting capture today holds the record from today forward and can never reconstruct what the source already replaced.

Every day of the subscription adds a day nobody can go back and acquire — which is the part of this that is not a software feature.

Designed to be read by a model, not just a person. The Sourceprint travels with the value, so an LLM or an agent working on the feed can state what a number is, where it came from and what not to conclude from it — without an engineer in the loop to supply the context.

Talk to us about scope and terms Questions and worked examples

One feed, every desk, whatever you build

It carries declared facts — statements market parties made about themselves in public filings, not conclusions we drew from them. Which is why it sits under whatever analysis you already run instead of competing with it.

Market access

Exposure by corporate parent. Site-of-care hierarchy. Channel structure.

Finance

The population behind an accrual. Channel exposure by entity class.

Legal

What was true on the date. Who held what during a period.

Compliance

Status on a transaction date. Registry agreement with the filing.

Commercial analytics

The rollup under the CRM, the forecast and the territory model.

Government affairs

Who is affected in this state, evidenced to a filing line.

Medical affairs

The legal entity behind a site. Research funding concentration.

Corporate development

What consolidated, when the record reflected it, what changed after.

The questions are yours. We supply the facts they rest on.

Every value carries its organisation, its meaning and its date — so whatever you build on it can account for itself when someone asks.

Methods run on the records

The preserved, Sourceprinted record is the product. Everything done with it, audit included, is a method: published, open, and run by you on your own data.

Every method asks one question of the records — was this true of this entity on this date? They differ only in when they ask it. Predicate Audit asks it of the past, transaction by transaction as of each governing date. Predicate Check asks it at the moment, as a transaction is made or a claim decided. Predicate Watch asks it as things change, raising an alert when an answer moves.

The methods are prebuilt, not prerun. We publish them and you run them in your own environment; no results come back to us and none go into the feed. Their code is open, so anyone — including the other side — can run them and get the same answer.

An audit escalates in three steps, and the records change what the first one can do

Each rung needs a different right of access, and each rung's output is written as the next one's exhibit. The rung that moves is the first. Status questions — was this entity, site or arrangement listed on the governing date — are settled outright at screening, against your own transactions, with no access to anyone else's records required.

HOW AN AUDIT ESCALATES
TierDataRight neededProduces What the records do
RUNG 1Screening Your own transactions, plus the independent dated record None Full-population flags, and a specific, dated showing Decide status questions outright
RUNG 2By agreement The counterparty's records under a contract or request Contract Findings backed by records Exhibit supporting reasonable cause or an audit demand
RUNG 3Compelled The counterparty's records under a regulator, statute or court Formal gate Enforceable findings Authenticated evidence
With LLM-assisted review, tier 1 can cover every transaction rather than a sample. The limits are no longer the cost of analysis but access to data and the time to review exceptions.
This is the commercial point of the feed, stated as a process. Most of what an audit wants to know used to require rung 2 or 3 — a contract, a regulator, a court. A dated record settles the status questions at rung 1, which is the only rung you can reach unilaterally, and does it across the whole population rather than a sample.

The guidelines below are where we have shown the properties working, in public, against primary sources.

340B eligibility

Whether an entity, a site or an arrangement was eligible on a given date — a question the public sources cannot answer, because they publish only their current state and overwrite the rest.

Read the 340B audit guideline →

Employer plans and PBMs

Testing pharmacy claims and rebates against the contract terms and public record in force on each date, rather than against today's.

Read the plan guideline →

Why the date is part of the fact

The principles behind the method, and the seven other domains where a status on a date decides an outcome and the source overwrites it.

Read the principles →

Predicate RX keeps the record. Methods use it.

Records are of the market. Methods run on the records.

Talk to us Questions and worked examples